How to Structure Warranty, RMA, and Spares for Reserved B300 Orders

A procurement team reserved a Supermicro HGX B300 deployment for a 24-month production program. Systems arrived in phases. Standard warranty on each lot began when that unit shipped. Six months later a power shelf failed. The replacement path required the failed shelf to return to a depot before a new one left. A nominal three-year hardware warranty existed on paper. The cluster still lost usable capacity inside the reserved window.

That is a support-commencement and restoration failure, not a missing logo on a warranty certificate. For reserved B300 purchase orders, coverage, RMA logistics, and spare-parts ownership should attach to the reserved capacity window and to commercially usable compute—not to a ship date that can start the clock before the system is even installed.

Start coverage when the system becomes operational

Shipment, arrival, installation, acceptance, and reserved-capacity commencement are different commercial facts. A warranty that begins at ship date can expire, or sit idle, out of alignment with the program that actually needs the machines. Write four dates, not one.

Warranty commencement is the date coverage becomes live for a named lot: typically installed and accepted, or another operational event both parties can witness. Reserved-capacity commencement is the date that lot is supposed to be commercially usable for the contracted class of work. Coverage expiration should be long enough to cover the reserved capacity window for that lot, including any agreed extension. A long-stop commencement is the latest date coverage can start if installation is delayed; without it, a late hall can leave hardware sitting uninsured, or start a manufacturer clock the buyer never intended to burn.

Treat phased lots separately. Lot 1 that shipped in March and lot 3 that ships in September should not share a single commencement sentence. Each lot needs its own warranty start, reserved-window coverage, and long-stop. A blended “three years from first shipment” clause is how later phases lose the back half of the program.

Name who provides the coverage: Pacific Intelligent Technologies, Inc., the manufacturer, an authorized service organization, or a combination with a written split. “OEM warranty included” is not a split. Write which party takes first response, which party owns depot logistics, and which party the buyer calls when those two disagree. Teams sourcing reserved GPU capacity from Pacific should treat support commencement as part of buying named compute, not as a datasheet footnote.

Put FRU response obligations in a support matrix

A warranty paragraph that says “next-business-day parts” without naming the field-replaceable unit will not restore a reserved cluster. Build a matrix. Rows are the parts that actually take a B300 environment offline. Columns are the response obligations you can administer.

• GPU baseboards and GPU modules.
• NVSwitch assemblies and related fabric silicon.
• CPUs, memory, and boot devices.
• NICs and host-fabric adapters.
• Power supplies and power shelves.
• Cooling parts, including CDUs, manifolds, and cold plates where they are FRUs.
• Fans and thermal or chassis controllers.
• Cables and optics.

For each row, specify fault-confirmation (who declares the FRU failed, and on what evidence), shipment of the replacement, delivery location, whether replacement happens on-site or after the failed unit leaves, diagnostics the buyer must run before an RMA is accepted, and the escalation path when the first responder cannot close the ticket. Measure restoration of usable reserved capacity, not ticket acknowledgement. An eight-hour ack on a power shelf that still sits in a depot is not a support obligation.

Keep the units in the matrix identical to the units on the purchase order. If the order reserves systems or usable B300 GPUs, restoration should be counted the same way. A replaced PSU that leaves the node unenumerated is not a restored FRU.

Choose depot, on-site, or advance replacement deliberately

Depot return, on-site swap, and advance replacement are different products. Define them by component. A cable or optic can often travel depot. A GPU baseboard or a power shelf on the reserved critical path usually cannot wait for the failed unit to arrive before the replacement ships.

Advance replacement means a good unit ships, or is drawn from a named spare pool, before or without waiting for the failed unit to reach a depot. On-site means a technician or a trained customer team installs the FRU at the reserved location. Depot means the failed unit must leave, be received, and often be diagnosed before a replacement is released. Mixing those three in one sentence is how a “next-day” clause becomes a two-week outage.

If a spare pool exists, write ownership, quantity, refresh, location, who can draw, and replenishment after use. A spare that the supplier still owns, stored two time zones away, that only the supplier can release after a ticket, is not a customer spare. A spare the customer owns but cannot refresh after the first draw is a one-time buffer, not a pool. Location should be specific enough that a night-shift operator can find the part without a new purchase order.

Quantity should be reasoned from the reserved window and the FRU classes in the matrix, not from a leftover crate. Refresh should say what happens when a spare is consumed: who ships the replacement spare, how long replenishment may take, and whether the consumed spare counts against warranty or against a separate parts account. Do not invent a rate card here. Write the control, not a price.

Align RMA logistics with capacity continuity

RMA paperwork is not the same as restored compute. Align the logistics with the reserved window: which FRUs are critical path, which can wait for depot, and what the buyer may run while a part is in transit.

Advance replacement should be the default for critical-path FRUs—GPU modules, NVSwitch, power shelves, and any cooling part that takes a reserved block offline. Depot-first is acceptable for non-critical parts if the contract says so. Silent “return first” language on a power shelf is how a 24-month program loses a month of usable capacity with a valid warranty still in force.

Write the documentation required to open and close an RMA: serials, fault codes, diagnostic artifacts, photos, packing requirements, and the address that actually receives the failed unit. Incomplete documentation should delay administration, not invent a new commencement date. If the supplier can reject an RMA for missing screenshots while the reserved cluster sits dark, those screenshots belong in the runbook before first production.

When a critical-path FRU will not restore inside the reserved window, treat interim compute as a separate commercial object. Pacific's bridge capacity is designed for the interval before longer-term capacity is usable again. It should not silently rewrite the warranty, the spare pool, or the reserved B300 baseline. State whether tendering interim capacity suspends support clocks, runs in parallel, or is simply cover while the RMA proceeds.

End-of-support and spares ownership

Warranty end, manufacturer end-of-support, and the last day of the reserved capacity window are three different dates. Write what happens when they diverge. If manufacturer support ends while the reserved program still has months to run, name who supplies FRUs, who may repair, and whether the spare pool transfers to the buyer.

Spares ownership at end of support should be explicit: which parts the buyer keeps, which the supplier retrieves, which must be sanitized or destroyed, and who pays freight. A spare pool that evaporates on the last warranty day leaves the reserved window unprotected even if the machines still run. A spare pool the buyer thought they owned, but that was only consigned, will be collected at the worst time.

Do not leave “reasonable commercial efforts to source parts after EOS” as the only sentence. Either the supplier has a parts obligation through the reserved window, the buyer owns a defined pool, or the program must plan a refresh. Ambiguity here is how a late-life power-shelf failure becomes an unplanned cluster retirement.

Pacific Intelligent Technologies, Inc. structures reserved capacity around the date compute becomes usable and stays usable. If your team is aligning warranty commencement, an FRU matrix, and spare-pool ownership with that same window, book 30 minutes with Harper.

FAQ

Should warranty start on the ship date?

Usually no, not as the only commencement. Shipment can be a progress milestone. Coverage for reserved B300 capacity should start when the lot is operational—typically installed and accepted—and should last through the reserved window. A long-stop commencement protects both parties if installation is delayed. Review Pacific's reserved GPU capacity offering when that distinction is the commercial issue.

Is a three-year OEM warranty enough for a 24-month reserved program?

Not by itself. A manufacturer term that starts at ship can be half-consumed before the reserved window even begins, and it says nothing about advance replacement, spare-pool ownership, or restoration of usable capacity. Align coverage expiration with the reserved window and write the FRU path separately.

Does depot RMA restore reserved capacity?

Only after the replacement is installed and the reserved quantity is usable again. Depot receipt of a failed power shelf is logistics, not restoration. For critical-path FRUs, prefer advance replacement or an on-site spare the buyer can actually draw.

Who should own the spare pool?

Whoever must restore capacity inside the reserved window without waiting for a new commercial release. Write ownership, quantity, location, who can draw, and replenishment. Consigned supplier inventory that cannot be touched after hours is not a buyer spare pool.

What if a critical FRU will not return during the reserved window?

Treat interim compute as cover, not as a silent rewrite of the reserved order. Bridge capacity can protect the workload while the RMA proceeds. Keep that cover on its own terms. When the warranty matrix, spare exhibit, and commencement dates are in one package, schedule 30 minutes with Harper.

The core rule is simple: start coverage when the reserved lot becomes operational, put FRU restoration in a matrix, choose depot versus advance replacement by component, own the spare pool you expect to draw, and write end-of-support before the last year of the program. That is how a three-year certificate stops masquerading as capacity continuity.